Multispectral images

Introduction

In this tutorial, you will learn how to import multispectral images to Supervisely using Python SDK and get the advantage of the grouped view in the labeling interface, which allows you to synchronize the view, zooming, panning, and labeling of images in one group.

You can also import multispectral images using Import Multispectral Images app from Supervisely Ecosystem.

You will learn how to:

For the advanced level, you can use supervisely annotation JSON format to download and upload projects with multispectral images.

Everything you need to reproduce this tutorial is on GitHub: source code and additional app files.

How to debug this tutorial

Step 1. Prepare ~/supervisely.env file with credentials. Learn more here.

Step 2. Clone the repository with source code and demo data and create a Virtual Environment.

git clone https://github.com/supervisely-ecosystem/import-multispectral-images-tutorial.git

cd import-multispectral-images-tutorial

sh create_venv.sh

Step 3. Open the repository directory in Visual Studio Code.

code -r .

Step 4. Change the Team ID and Workspace ID in the main.py file by copying the ID from the context menu.

team_id = 448
workspace_id = 690

Step 5. Start debugging src/main.py.

Supervisely instance version >= 6.8.54 Supervisely SDK version >= 6.72.201\

In the tutorial, Supervisely Python SDK version is not directly defined in the requirements.txt. But when developing your app, we recommend defining the SDK version in the requirements.txt.

Import libraries

import cv2
import os
import tifffile
import supervisely as sly
import nrrd
from dotenv import load_dotenv

Enter your Team ID and Workspace ID

team_id = 448
workspace_id = 690

Init API client

load_dotenv(os.path.expanduser("~/supervisely.env"))
api = sly.Api.from_env()

Create a new project and dataset

project = api.project.create(workspace_id, "Multispectral images", change_name_if_conflict=True)
dataset = api.dataset.create(project.id, "ds0")

Set multispectral settings for the project

To enable grouping of images, including view, zooming, panning and labeling, you need to set multispectral settings for the project. You can do it with just one line of code:

api.project.set_multispectral_settings(project.id)

And now we're ready to upload images.

How to upload multispectral images

In this tutorial, we'll be using the api.image.upload_multispectral method to upload images to Supervisely.

def upload_multispectral(
        dataset_id: int,
        image_name: str,
        channels: Optional[List[np.ndarray]] = None,
        rgb_images: Optional[List[str]] = None,
    ) -> List[ImageInfo]:

Parameters	Type	Description
dataset_id	int	ID of the dataset to upload
image_name	str	Name of the image will be used as name of the images group
channels	Optional[List[np.ndarray]] = None	List of channels as 2d numpy arrays.
rgb_images	Optional[List[str]] = None	List of paths to RGB images.

Parameters

Type

Description

dataset_id

int

ID of the dataset to upload

image_name

str

Name of the image will be used as name of the images group

channels

Optional[List[np.ndarray]] = None

List of channels as 2d numpy arrays.

rgb_images

Optional[List[str]] = None

List of paths to RGB images.

So, the method uploads images (which can be passed as channels or RGB images) to Supervisely and returns a list of ImageInfo objects. RGB images as paths or channels as NumPy arrays can be passed to the method or both at the same time. The result will be a group of images in both cases.

Below, you will find examples of different ways to load multispectral data. However, it is important to note that you can group the images in any way you like: split them into channels, load them as a whole, or both at the same time. It is entirely up to you.

Upload an image as channels

Input: 1 RGB image in PNG format. Output: 3 images in a group with the name demo1.png in Supervisely.\

image_name = "demo1.png"
image = cv2.imread(f"demo_data/{image_name}")

# Extract channels as 2d numpy arrays: channels = [a, b, c]
channels = [image[:, :, i] for i in range(image.shape[2])]

image_infos = api.image.upload_multispectral(dataset.id, image_name, channels)

We'll do this operation for other cases too, so let's take a closer look at the code:

The image_name will be used as a group name in the labeling interface and it can be any string.
Then we read the image from the disk using OpenCV.
Now we're splitting the image into channels, and preparing a list of channels as NumPy arrays.
Finally, we upload the channels to Supervisely using the api.image.upload_multispectral method.
The method returns a list of ImageInfo objects, which contain information about the uploaded images. Learn more about ImageInfo here.

So, those are the steps we'll be doing for all the other cases.

Upload a multichannel tiff image as channels

Input: 1 multichannel tiff image. Output: 7 images in a group with the name demo2.tif in Supervisely.\

image_name = "demo2.tif"
image = tifffile.imread(f"demo_data/{image_name}")

# Extract channels as 2d numpy arrays: channels = [a, b, c, d, e, f]
channels = [image[:, :, i] for i in range(image.shape[2])]

image_infos = api.image.upload_multispectral(dataset.id, image_name, channels)

Upload nrrd image as channels

In this tutorial, we'll be uploading the channels of nrrd image as separate images. But Supervisely supports high-dimensional nrrd images, so you can upload them as is without splitting them into channels.

Input: 1 nrrd image. Output: 7 images in a group with the name demo3.nrrd in Supervisely.\

image_name = "demo3.nrrd"
image, header = nrrd.read(f"demo_data/{image_name}")

# Extract channels as 2d numpy arrays: channels = [a, b, c, d, e, f]
channels = [image[:, :, i] for i in range(image.shape[2])]

image_infos = api.image.upload_multispectral(dataset.id, image_name, channels)

Upload a pair of RGB and thermal images without splitting them into channels

Input: 1 RGB image and 1 thermal image in PNG format. Output: 2 images in a group with the name demo4.png in Supervisely.\

image_name = "demo4.png"
images = ["demo_data/demo4-rgb.png", "demo_data/demo4-thermal.png"]

image_infos = api.image.upload_multispectral(dataset.id, image_name, rgb_images=images)

As you can see, in this case, we don't extract any channels since we need to upload only images, not channels. So, we pass the list of image paths to the rgb_images parameter.

Upload RGB image, its channels, and depth image

Input: 1 RGB image and 1 depth image in PNG format. Output: 5 images in a group with the name demo5.png in Supervisely.\

image_name = "demo5.png"
images = ["demo_data/demo5-rgb.png", "demo_data/demo5-depths.png"]

image = cv2.imread(images[0])

# Extract channels as 2d numpy arrays: channels = [a, b, c]
channels = [image[:, :, i] for i in range(image.shape[2])]

image_infos = api.image.upload_multispectral(dataset.id, image_name, channels, images)

Here, we uploaded one image both as channels and as an image. So, we pass the list of image paths to the rgb_images parameter and the list of channels to the channels parameter.

Upload grayscale and UV images

Input: 1 grayscale image and 1 UV image in PNG format. Output: 2 images in a group with the name demo6.png in Supervisely.\

image_name = "demo6.png"
images = ["demo_data/demo6-grayscale.png", "demo_data/demo6-uv.png"]

image_infos = api.image.upload_multispectral(dataset.id, image_name, rgb_images=images)

Upload RGB image, thermal image, and channels of thermal image

Input: 1 RGB image and 1 thermal image in PNG format. Output: 5 images in a group with the name demo7.png in Supervisely.\

image_name = "demo7.png"
images = ["demo_data/demo7-rgb.png", "demo_data/demo7-thermal.png"]

image = cv2.imread(images[1])

# Extract channels as 2d numpy arrays: channels = [a, b, c]
channels = [image[:, :, i] for i in range(image.shape[2])]

image_infos = api.image.upload_multispectral(dataset.id, image_name, channels, images)

Upload RGB image and two MRI images

Input: 1 RGB image and 2 MRI images in PNG format. Output: 3 images in a group with the name demo8.png in Supervisely.\

image_name = "demo8.png"
images = ["demo_data/demo8-rgb.png", "demo_data/demo8-mri1.png", "demo_data/demo8-mri2.png"]

image_infos = api.image.upload_multispectral(dataset.id, image_name, rgb_images=images)

Grouped view in the labeling interface

So now, that we've uploaded all the images, let's take a look at the labeling interface.

As you can see, all the images are grouped by the name of the group, which is the name of the image we passed to the image_name parameter. We can zoom, pan, and label images in one group at the same time. So, whenever you create a label on one image, it will be automatically created on all the other images in the group. You can edit label on another image in the group, and it will be automatically updated on all the other images in the group. Just a reminder: we set the multispectral settings for the project at the beginning of the tutorial with the api.project.set_multispectral_settings method, which enables this grouped view.

Summary

In this tutorial, you learned how to upload multispectral images to Supervisely using Python SDK and get the advantage of the grouped view in the labeling interface, which allows you to synchronize the view, zooming, panning, and labeling of images in one group. Let's recap the steps we did:

Create a new project and dataset.
Set multispectral settings for the project using the api.project.set_multispectral_settings method.
Upload images using the api.image.upload_multispectral method.

And that's it! Now you can upload your multispectral images to Supervisely using Python SDK.

(Advanced) Use supervisely format for multispectral images

You can always use Export to Supervisely format to download to the local directory of your favorite multispectral project with the preserved multi-view settings and then easily upload it as a new project to the Import images in Supervisely format

From the developer's point of view, the Supervisely annotation JSON format gives you easy access to the necessary parameters while grouping the images. To feel the power of this instrument, let's imagine the situation when you have already downloaded the project and opened the meta.json file:

{
  "classes": [
    {
      "title": "leaf",
      "shape": "bitmap",
      "color": "#FB00ED",
      "geometry_config": {},
      "id": 6509759,
      "hotkey": ""
    }
  ],
  "tags": [
    {
      "name": "im_id",
      "value_type": "any_number",
      "color": "#0F8A2D",
      "id": 27855,
      "hotkey": "",
      "applicable_type": "all",
      "classes": []
    },
    {
      "name": "band",
      "value_type": "any_string",
      "color": "#9F7A2F",
      "id": 27856,
      "hotkey": "",
      "applicable_type": "all",
      "classes": []
    }    
  ],
  "projectType": "images",
  "projectSettings": {
    "multiView": {
      "enabled": true,
      "tagName": "im_id",
      "tagId": 27855,
      "isSynced": false
    }
  }
}

This is very easy to see, if you want to group your images by the band tag instead of im_id (sync mode on), simply change your projectSettings this way:

  "projectSettings": {
    "multiView": {
      "enabled": true,
      "tagName": "band",
      "tagId": 27856,
      "isSynced": true
    }
  }

Moreover, you can additionally enhance your tags with hotkey, or specify the tag classes. See the explanation of every field here.

To download and upload a project using Supervisely SDK, use the following code:

import supervisely as sly
from tqdm import tqdm

load_dotenv(os.path.expanduser("~/supervisely.env"))
api = sly.Api.from_env()
# id of your workspace and project from environmental variables
# see https://developer.supervisely.com/getting-started/environment-variables 
WORKSPACE_ID = sly.env.workspace_id()
PROJECT_ID = sly.env.project_id()

your_dir = "/your/multi-view/project/dir"

project_info = api.project.get_info_by_id(PROJECT_ID)
pbar = tqdm(desc="Download Project", total=project_info.items_count)
sly.download_project(api, PROJECT_ID, your_dir, progress_cb=pbar)

project_fs = sly.read_project(your_dir)
pbar = tqdm(desc="Upload Project", total=project_fs.total_items)
sly.upload_project(project_fs.directory, api, WORKSPACE_ID, progress_cb=pbar)

PreviousKeypoints (skeletons)NextMulti-view images

Last updated 2 days ago